Cong Hu

Full-duplex (FD) communication enables simultaneous transmission and reception on the same frequency band; however, considerable self-interference (SI) constrains the achievable signal-to-interference-plus-noise ratio (SINR) and overall system performance. This study examines an FD scenario where the communication link between the base station and a remote terminal is exclusively facilitated by reconfigurable intelligent surfaces (RIS), without a direct line-of-sight path. To address the SI challenge, we introduce a novel optimization algorithm, Dual-Memory Neural Population Dynamics (DMANPD), which utilizes a dual-memory framework integrating long-term and short-term memory mechanisms. Long-term memory facilitates global exploration through novelty and competition, while short-term memory focuses on local refinement strategies to adapt to the changing search environment. Simulation results from the IEEE Congress on Evolutionary Computation 2017 (CEC 2017) benchmark suite indicate that DMANPD achieves an average fitness superiority rate of 79.31%. In RIS-assisted FD systems, the proposed scheme consistently achieves higher received SINR compared to baseline methods, with a minimum received SINR surpassing benchmark algorithms by approximately 5-10 dB across various SNR regimes, indicating robustness under diverse operating conditions. Furthermore, for large-scale RIS arrays, DMANPD reduces the per-iteration time by more than 15 ms and shortens the convergence time to achieve the 20 dB SINR threshold— necessary for high-speed FD communication—by over 1 s, demonstrating significant computational efficiency gains.

Radio Frequency Fingerprint Identification (RFFI) is a technology that utilizes unique hardware characteristics for identity authentication. This paper proposes a feature fusion-based RFFI approach employing an enhanced residual network (ResNet) to improve recognition accuracy under various signal-to-noise ratio (SNR) conditions. First, pulse sequences are extracted, and the following features are then constructed: ambiguity function (AF), time-frequency spectrograms (TFS), bispectral analysis (BIS), and intrinsic function-phase feature (IFP). The network architecture is based on an enhanced ResNet with a multi-path structure, gated recurrent unit (GRU), and attention mechanisms. We ultimately validate our approach using real-world automatic dependent surveillance-broadcast (ADS-B) data. Experimental results reveal the following key findings: 1) For the original signal with an SNR of 25 dB, the accuracy of the method stacking four features improves by 4.55% compared to the method that stacks two features. 2) By leveraging the AF feature, a notable accuracy gain of 24.3% is achieved by the enhanced ResNet compared to the classical ResNet-34. 3) Employing a combination of four features with the enhanced ResNet, this approach yields a 31.81% accuracy improvement over the method that uses AF with ResNet-34. 4) With knowledge distillation applied, the ShuffleNet-V2, using ResNet-G as the teacher model for distillation, maintains a parameter size of only 2.28M, while suffering a mere 1.33% accuracy loss.

In wireless communications, radio frequency fingerprint identification (RFFI) leverages unique hardware characteristics for device recognition. This paper proposes an innovative few-shot RFFI method based on adaptive energy path integrated projection (AEP-IP) and an enhanced neural network, aiming to significantly improve recognition accuracy under various signal-to-noise ratio (SNR) conditions. The method integrates an enhanced densely connected temporal convolutional network (DenseTCN), transfer learning, and Meta-Learning (MAML). Key steps include extracting pulse sequences from received signals and constructing four features: ambiguity function (AF), time-frequency spectrogram (TFS), bispectrum analysis (BIS), and intrinsic function-phase (IFP). One-dimensional projection features are extracted through these features. The DenseTCN enables cross-layer feature fusion and serves as the backbone of the Trans-MAML scheme to enhance few-shot learning under low SNR conditions. Validated using real-world automatic dependent surveillance-broadcast (ADS-B) data, the results demonstrate: (1) Four-timestep feature improve accuracy by 7.9% (25 dB SNR) and 11% (0 dB SNR) compared to two-timestep feature. (2) DenseTCN outperforms TCN by 8.86% under the same conditions. (3) DenseTCN achieves comparable performance to TCN with 100% data using only 20% data (25 dB SNR). (4) With limited training data, the Trans-MAML method achieves higher recognition accuracy across all SNR levels compared to baseline methods.

Full-duplex (FD) systems enable users to simultaneously transmit and receive information on the same frequency, theoretically enhancing spectrum efficiency. However, the inherent openness and broadcasting nature of wireless communication pose significant security risks, especially in the presence of eavesdroppers. To enhance the reliability of FD relay systems, we derive the secrecy capacity of the systems and propose a novel approach that integrates relay selection and power allocation utilizing deep reinforcement learning (DRL). Specifically, our method employs deep Q network (DQN) and proximal policy optimization (PPO) algorithms to jointly explore the optimal relay selection and power allocation strategies. Simulation results indicate that the developed approach enhances the average secrecy capacity by at least 30% compared to benchmark methods. Additionally, we assessed the performance of various methods under diverse conditions, demonstrating the effectiveness and adaptability of the proposed approach.

In this article, we propose a multi-impairment compensation scheme based on deep learning and low-density parity-check (LDPC) code for the coherent optical orthogonal frequency division multiplexing (CO-OFDM) system. We first propose a multi-impairment compensation autoencoder (MICAE) based on deep neural network (DNN). Then we combine the proposed MICAE with LDPC code and design a DNN-based decoder for LDPC decoding to replace the traditional belief propagation (BP) decoder. The proposed scheme can compensate for multiple impairments simultaneously and has faster decoding speed, greatly improving the performance of the CO-OFDM transmission system. We demonstrate the superiority of the proposed scheme through simulations in different CO-OFDM transmission systems. Simulation results show that the proposed scheme can effectively improve the Q-factor and reduce the bit error rate (BER) of the system, and is more suitable for complex long-distance and high-speed optical transmission scenarios.

Automatic Modulation Classification (AMC) is a critical technology in the field of non-cooperative communication. Carrier frequency offset is an inevitable impairment factor in communication systems, and it significantly affects the performance of AMC, especially when the range is large. Therefore, frequency offset estimation and compensation are essential for automatic modulation classification. To address this issue, this paper proposes an innovative cascading network based on deep learning, primarily comprising pre-classification and compensation mechanisms. With the pre-classification mechanism, we categorize the eleven unknown modulation formats into three major classes: PSK, FSK, and QAM. Subsequently, we utilize deep neural networks to estimate the frequency offset of signals within these major classes, followed by performing the corresponding signal compensation. The experimental results demonstrate that the pre-classification mechanism effectively reduces the error in frequency offset estimation, minimizing the error (mean-square error) to as low as -3.9. Under various signal-to-noise ratio (SNR) ranges, the classification accuracy achieved through pre-classification and compensation is up to 14% higher than the accuracy without this method. Specifically, at SNR = 10 dB, it can reach 99%. Moreover, the proposed scheme is robust and performs stably across different frequency offsets range, outperforming other existing network classification schemes to some extent.

In the context of deep-space wireless communication, the complexity of the cosmic environment often leads to the degradation of received signal quality. In such intricate environments, accurately assessing the signal-to-noise ratio (SNR) becomes paramount, as it directly impacts the reliability, performance, and success rate of data transmission within the communication system. However, traditional SNR estimation methods have limitations, such as accuracy issues and poor adaptability, especially in low SNR scenarios. Given the widespread adoption of deep learning methods in various domains, this paper introduces a novel SNR estimation technique. This approach leverages a network architecture that combines elements of Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM) algorithms and attention mechanism, constituting a Non-Data Assisted (NDA) methodology. To evaluate the effectiveness of our technique, we construct a simulated QPSK dataset with varying SNR levels in deep space channels and conduct a series of simulation experiments. These experiments involve a detailed comparison of our proposed approach with existing algorithms, including the Second and Fourth Moments (M2M4) algorithm, the Signal-to-Noise Variance (SNV) algorithm, the Signal-to-Variation Ratio (SVR) algorithm, and the Singular Value Decomposition (SVD) algorithm. Additionally, we conducted an exhaustive comparative analysis with classical neural network methodologies. Empirical findings undeniably establish the superiority of our proposed approach, outperforming both traditional methods and classical deep learning techniques, showcasing its superior estimation capabilities.

Efficient optical network management poses significant importance in backhaul and access network communication for preventing service disruptions and ensuring Quality of Service (QoS) satisfaction. The emerging faults in optical networks introduce challenges that can jeopardize the network with a variety of faults. The existing literature witnessed various partial or inadequate solutions. On the other hand, Machine Learning (ML) has revolutionized as a promising technique for fault detection and prevention. Unlike traditional fault management systems, this research has three-fold contributions. First, this research leverages the ML and Deep Learning (DL) multi-classification system and evaluates their accuracy in detecting six distinct fault types, including fiber cut, fiber eavesdropping, splicing, bad connector, bending, and PC connector. Secondly, this paper assesses the classification delay of each classification algorithm. Finally, this work proposes a fiber optics fault prevention algorithm that determines to mitigate the faults accordingly. This work utilized a publicly available fiber optics dataset named OTDR_Data and applied different ML classifiers, such as Gaussian Naive Bayes (GNB), Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), and Decision Tree (DT). Moreover, Ensemble Learning (EL) techniques are applied to evaluate the accuracy of various classifiers. In addition, this work evaluated the performance of DL-based Convolutional Neural Network and Long-Short Term Memory (CNN-LSTM) hybrid classifier. The findings reveal that the CNN-LSTM hybrid technique achieved the highest accuracy of 99% with a delay of 360 s. On the other hand, EL techniques improved the accuracy in detecting fiber optic faults. Thus, this research comprehensively assesses accuracy and delay metrics for various classifiers and proposes the most efficient attack detection system in fiber optics.